From Tools to Agents: What Has Actually Changed

For most of the history of computing, software has waited to be told what to do. Even the most sophisticated systems answered a question, produced a forecast or flagged an exception, and then stopped. A person read the output and decided what happened next.

That boundary is moving. The newest systems can take a goal, break it into steps, use other software to carry those steps out, check the results and continue, sometimes across dozens of actions before a person sees anything at all. This is the shift I call the Age of Agency: from AI that answers to AI that acts.

It is easy to overstate this change, and just as easy to dismiss it. This essay sets out what is genuinely new, what is merely relabelled, and why the difference matters to anyone who must decide how far to trust these systems.

The boundary that moved

The important change is not that machines have become cleverer, although they have. It is that the point at which a person intervenes has moved.

In a traditional system, the human sits between every output and every consequence. A credit model produces a score; an underwriter approves the loan. A forecasting tool predicts demand; a planner places the order. The machine informs. The person acts.

In an agentic system, the human moves upstream. A person sets the goal and the limits, and the system decides the steps and takes them. The underwriter becomes the author of a policy the agent applies. The planner becomes the supervisor of an agent that places orders within agreed bounds. The machine does not just inform the decision. It carries it out.

That is why the step from recommending to acting is the one that matters. A wrong recommendation can be caught. A wrong action has already happened.

A spectrum of autonomy

Much confusion comes from treating "AI" as one thing. It is more useful to think of a spectrum with three broad positions.

Position	What the system does	Where the person sits
Assistance	Drafts, summarises, suggests	Reviews every output before anything happens
Automation	Executes a fixed procedure designed by people	Designs the procedure and handles exceptions
Agency	Pursues a goal, chooses its own steps, uses tools and acts	Sets goals and limits, supervises results

Automation is not new. Organisations have automated procedures for decades, and the rules were written by people in advance. What distinguishes agency is that the system itself chooses the route. It can respond to situations nobody anticipated, which is both its power and its risk.

Most real systems today sit somewhere between these positions, and many move along the spectrum over time as they earn trust. The point is not to classify every tool precisely, but to ask the right question about each one: at what point, if any, does a person decide before something happens?

Why acting changes the economics, the risks and the management

Three things change when a system moves from recommending to acting.

The economics change. When a person must act on every output, the cost of each decision includes their time. When the system acts, the marginal cost of a decision falls towards the cost of computation. Work that was never worth doing at human cost becomes worth doing. Activity can scale without headcount scaling with it. This is the source of the large productivity claims made for agents, and it is real, but it only appears when the system is trusted to act.

The risks change. A person reviewing outputs one at a time is slow, but slowness is also a safeguard. An agent can repeat the same error across thousands of cases before anyone notices. It can act on manipulated instructions hidden in an email or a document. And because it chooses its own steps, it can reach outcomes that nobody designed. The risk is no longer only that the model is wrong. It is that nobody can say who allowed it to act.

The management changes. Buying a tool is a procurement decision. Deploying an agent is a delegation decision, closer to appointing a member of staff than to installing software. It raises the questions organisations ask about any delegate: what may it do, within what limits, who supervises it, and how do we know it is behaving as intended?

What is genuinely new, and what is relabelled

The word "agent" is now applied to almost everything. A chatbot with a new name is not an agent. Nor is a scripted workflow with a language model attached to one step.

Four questions separate the genuine article from the relabelled one:

1. Does the system pursue a goal, rather than respond to a single request?
2. Does it choose its own steps, rather than follow a fixed script?
3. Does it use other tools and systems, such as databases, email or payment platforms?
4. Does it act without a person approving each step?

A system that answers yes to all four is an agent in the sense that matters for governance. A system that answers no to the last question is still assistance, however impressive. Boards should ask these questions of every proposal that uses the word.

Three strands that are often blurred

Discussion of agents tends to mix three different developments.

Digital agents work entirely in software: reconciling accounts, handling service requests, preparing documents, moving data between systems. They are spreading fastest because the environment they act in is already digital and the cost of an error is often recoverable.

Physical systems act in the world: robots in warehouses, autonomous vehicles, systems that control energy grids or building plant. Their errors can cause physical harm, so they are governed by safety disciplines that digital agents have largely escaped.

The convergence of the two is where the most consequential questions will arise: digital agents that instruct physical systems, such as an agent that schedules maintenance and then dispatches a robot to perform it. Here, a software error becomes a physical event.

Treating these as one category leads to poor decisions. The controls appropriate for an agent that drafts letters are not those appropriate for one that moves money, and neither is sufficient for one that moves machinery.

A shared vocabulary for the board

The aim of Part I of The Age of Agency is a common language, so that the decisions which follow are made on what a system does, not on what it is called. In practice, that means every board and executive team should be able to answer five questions about any agent in use or proposed:

1. Where does it sit on the spectrum: assistance, automation or agency?
2. What can it act on: which systems, data, money or physical equipment?
3. At what point, if any, does a person decide before it acts?
4. What happens when it is wrong, and how quickly would we know?
5. Who in the organisation answers for what it does?

These questions do not require technical expertise. They require clarity about delegation, which is a subject boards already understand.

What follows

Once it is clear that agents act, the next question is unavoidable: which decisions should they be allowed to make? That is the subject of Part II, and of the book's central tool, the delegation map.

This essay introduces Part I of the forthcoming book The Age of Agency. The ideas will be developed in full in the chapters that follow.

Views are the author's own. The frameworks described are applied methodologies of Kiros Intelligence Group, not regulatory standards.